OpenEvidence MCP (Unofficial)

OpenEvidence MCP is an unofficial Model Context Protocol server that connects OpenEvidence to Codex, Claude Code, Claude Desktop, Cursor, Cline, Continue, and other MCP-compatible clients through your own authenticated browser session.

Translations: Русский | Español | 简体中文 | 繁體中文（台灣） | 한국어 | हिन्दी

Copy/Paste Agent Install Prompt

Using Codex, Claude Code, or another local AI coding agent? Copy this prompt into the agent and let it handle setup, MCP config, login guidance, and verification.

Look into this repository: https://github.com/bakhtiersizhaev/openevidence-mcp

Install OpenEvidence MCP in my local AI CLI / agentic MCP setup. Add it as an MCP server for the CLI or app I am using. Follow the repository README and the agent install playbook at docs/AGENT_INSTALL_PROMPT.md.

Verify local prerequisites: Node.js 20+, npm, git, and Playwright Chromium. Clone or update the repo, run npm ci, npx playwright install chromium, npm run build, and npm run check.

Configure the MCP server with command "node" and args pointing to the absolute path of dist/server.js. Keep the server local and do not expose it over a public network.

Guide me through OpenEvidence login with my own account. First try npm run login. If Google says "This browser or app may not be secure", stop that flow and run npm run login:browser instead. I will complete login in the opened browser window and then press Enter in the terminal.

Do not ask for or expose my password, cookies, tokens, storage-state files, screenshots with private account data, patient-identifiable information, or account identifiers. Do not bypass OpenEvidence, Google, institutional, regional, or account access controls.

After login, run npm run smoke. If smoke returns ok: true and authenticated: true, show me the final MCP config and tell me to restart my AI agent / MCP client so the OpenEvidence tools become available.

Longer agent runbook: docs/AGENT_INSTALL_PROMPT.md.

What it does

OpenEvidence MCP runs a local stdio MCP server that lets MCP clients use your existing OpenEvidence browser session for:

• checking whether the saved session is authenticated;

• listing your OpenEvidence question/article history;

• fetching a full article payload by ID;

• asking an OpenEvidence research question and optionally waiting for completion;

• polling an existing OpenEvidence article until it completes.

No official OpenEvidence API token is required.

What it does NOT do

• It is not affiliated with, endorsed by, or approved by OpenEvidence.

• It does not provide medical advice or replace clinical judgment.

• It does not bypass authentication, paywalls, or access controls.

• It does not collect credentials.

• It does not send your browser session state anywhere except to OpenEvidence through local Playwright requests.

• It should not be used for patient-specific diagnosis or treatment decisions without appropriate human review.

Who it is for

• clinicians using their own OpenEvidence account;

• medical researchers;

• AI operators building evidence-research workflows;

• MCP developers integrating local tools with Codex, Claude, Cursor, Cline, Continue, or similar clients.

Tested / Target Clients

This project is designed for MCP-compatible clients and local agent workflows. Only Codex and Claude-style local configuration examples are maintained in this repository unless otherwise noted.

Other MCP-compatible hosts may work as well, but the examples in this repository focus on Codex and Claude-style local MCP configuration.

Features

Agent Tool-Calling Notes

The MCP server includes built-in instructions and a prompt named openevidence_research_workflow for clients that expose MCP prompts.

Recommended agent workflow:

1. Call oe_auth_status when auth state is unknown.

2. Use oe_history_list only when the user wants prior OpenEvidence work or an article ID.

3. Use oe_article_get when you already have an article ID.

4. For long research questions, call oe_ask with wait_for_completion=false, then call oe_article_wait with the returned article_id.

5. Use original_article_id only for true follow-up continuity. Omit it for fresh questions to avoid stale thread context.

6. Treat outputs as evidence-research context, not medical advice, diagnosis, or clinical orders.

Related commands:

Requirements

• Node.js 20+

• npm 10+

• OpenEvidence account

• macOS, Windows, or Linux

• Chromium installed by Playwright (npx playwright install chromium)

Availability Note

OpenEvidence availability may depend on region, account eligibility, and OpenEvidence policy. Public materials in May 2026 indicate verified U.S. HCP/NPI-centered access and EU/U.K. unavailability; this project does not bypass those restrictions.

Useful references:

• OpenEvidence homepage

• OpenEvidence API/product page

• OpenEvidence Privacy Policy

Quick Start

macOS

git clone https://github.com/bakhtiersizhaev/openevidence-mcp.git
cd openevidence-mcp
./scripts/setup-macos.sh
npm run login
npm run smoke

Ubuntu/Linux

git clone https://github.com/bakhtiersizhaev/openevidence-mcp.git
cd openevidence-mcp
./scripts/setup-ubuntu.sh
npm run login
npm run smoke

Windows PowerShell

git clone https://github.com/bakhtiersizhaev/openevidence-mcp.git
cd openevidence-mcp
.\scripts\setup-windows.ps1
npm run login
npm run smoke

Login Flow

Run:

npm run login

The command opens a browser window. Sign in to OpenEvidence with your own account, return to the terminal, and press Enter. The login script validates /api/auth/me and saves local browser session state.

Default state paths:

• macOS/Linux: ~/.openevidence-mcp/auth/storage-state.json

• Windows: %USERPROFILE%\.openevidence-mcp\auth\storage-state.json

You can import an existing Playwright storage state file:

npm run login -- --import /absolute/path/storage-state.json

If Google sign-in says the browser or app may not be secure, use the system-browser login flow:

npm run login:browser

This opens Chrome or Edge with a local OpenEvidence MCP profile. Complete login in that browser, return to the terminal, and press Enter. The script saves local session state and verifies /api/auth/me.

Do not share storage-state files, cookies, screenshots with private account data, or patient-identifiable information.

MCP Client Setup

Build before registering the server:

npm run build

Codex

Add this to ~/.codex/config.toml:

[mcp_servers.openevidence]
command = "node"
args = ["/ABSOLUTE/PATH/openevidence-mcp/dist/server.js"]
startup_timeout_sec = 60

Windows example:

[mcp_servers.openevidence]
command = "node"
args = ["C:\\Users\\<user>\\openevidence-mcp\\dist\\server.js"]
startup_timeout_sec = 60

Claude Desktop

Add this to claude_desktop_config.json:

{
  "mcpServers": {
    "openevidence": {
      "command": "node",
      "args": ["/ABSOLUTE/PATH/openevidence-mcp/dist/server.js"]
    }
  }
}

Cursor, Cline, Continue

Use the same stdio server shape if your client supports MCP server command/args configuration:

{
  "command": "node",
  "args": ["/ABSOLUTE/PATH/openevidence-mcp/dist/server.js"]
}

Example configs are in examples/.

Verify

npm run smoke

Expected result with a valid session:

• ok: true

• authenticated: true

• a redacted history preview

If smoke fails with an auth error, run npm run login again. Smoke requires a real OpenEvidence account session and will not pass in a clean CI environment unless session state is provided securely.

By default, smoke output redacts account and history content. Use npm run smoke -- --verbose only in a private terminal if raw account/history payloads are needed for debugging.

Developer checks:

npm test
npm run build
npm run check

Security Notes

• Treat storage-state.json, cookies, and browser profiles as secrets.

• Do not commit .env, session state, screenshots with account data, or patient-identifiable information.

• Use only your own OpenEvidence account.

• Keep MCP client configs pointed at the built local server path you control.

• Review tool calls from autonomous agents before using outputs in clinical or operational workflows.

• See SECURITY.md for vulnerability reporting and supported scope.

Troubleshooting

See docs/TROUBLESHOOTING.md for detailed recovery steps.

Common fixes:

• authenticated: false: rerun npm run login.

• Google says browser is not secure: run npm run login:browser.

• Browser install errors: run npx playwright install chromium.

• MCP client cannot start server: confirm npm run build succeeded and use an absolute path to dist/server.js.

• Windows path issues: escape backslashes in JSON/TOML or use full absolute paths.

• Node errors: confirm node --version is 20 or newer.

• OpenEvidence UI/API changed: open an issue with sanitized logs and no private account or patient data.

Roadmap

• Keep tool descriptions compact and agent-friendly.

• Add focused tests around config and response parsing.

• Improve smoke diagnostics without exposing session details.

• Track MCP client setup examples as client configuration formats evolve.

License & Attribution

Apache-2.0 (LICENSE) + NOTICE.

If you redistribute, fork, or build derivative versions, keep attribution to:

• Original author: Bakhtier Sizhaev

• Original repository: https://github.com/bakhtiersizhaev/openevidence-mcp

Suggested attribution line:

Based on OpenEvidence MCP by Bakhtier Sizhaev - https://github.com/bakhtiersizhaev/openevidence-mcp